KrystalCWT

我的也是loss一直降不下去，请问你解决了吗？

请问你解决了吗？我也发现直接用yolo fast无法处理batch数据

> > 请问你解决了吗？我也发现直接用yolo fast无法处理batch数据 > > 支持批没多大意义。支持批的话，由于不同图片检测出的bbox数目不同，需要解决对齐问题。而且，批量后处理只是用于评测，实际应用时还是批大小为1. 那请问如果我yolo是batch predict的，那在fast yolo里面需要怎么修改nms部分呢？ 因为我现在希望把后处理部分也打包在网络里

> I have resolved this problem. I need each region feature, you can find in lib/model/faster_rcnn/faster_rcnn.py (line 86): pooled_feat = self._head_to_tail(pooled_feat). I need pooled_feat, if I use resnet-101 model, pooled_feat...

Hi, Can you share the pretrained model with baidu link? Thanks

> Note that cache files *.pkl (found under ~/data/COCO/cache, for e.g. the COCO dataset) must be re-generated. Otherwise, the code will load that cache file. Seems like an obvious thing,...

> > I have the same problem, how do you deal with this issue > > I have solve this problem, trained more time and change the lr. I directly...

> Ps o I have the same problem. Did you solve it?

same issue. I want to change the class num, but it does not work.

Have you figured out what the problem might be? Having nearly the same precision as yours... @cosimo17